ExampleColorHistogram run wild

[Bill Ross]

I just found I can get to about 75% accuracy in predicting interesting photo pairs using just greyscale:128 and saturation:128 uncorrelated histograms (S from HSV) - i.e. 256 numbers. It trains on 750K cases in ~30 sec, and predicts 11K^2 pairs in ~25 min on an nvidia 2070. I just spiked the ball here:

https://www.reddit.com/r/learnmachinelearning/comments/hokdlz/weekly_showoff/

I've grown ExampleColorHistogram into a sort of swiss army knife for a variety of purposes, and I'm thinking of making it the lead item in open sourcing my own code. I'm rusty on git, but if it's wanted in the boofcv tree, I'm glad to supply it one way or another.

Added user features include:

  - Give it a dir and a histogram type/dimension to use, and it histos all pics and then lets you search with an example pic - expansion of the match demo

  - Generate histos in bulk for ML or faster load for above option

  - Calc distance between histos using various algs (Euclid, Poincare spherical, another; earth-mover's would be nice to have)

E.g. the cmdline API for specifying a table name_type for distance between photo greyscale(128) distributions: b4_grey-gs_128-dist

And for specifying an angle measurement in an rgb 2x2x2 histogram space: d_rgb2-rgb_2-angle

Anyone interested?

Bill

Here you can see some other photo pairs through the ages:

https://www.linkedin.com/in/bill-ross-phobrain/detail/recent-activity/shares/

How to say 2020 beat us? 2021.

[Bill Ross]

And it's multithreaded.